Document 11070440

advertisement
LIBRARY
OF THE
MASSACHUSETTS INSTITUTE
OF TECHNOLOGY
Dewey
KD2.E
.M414
FEB
.U^'li
6
1974
M.a c 3.
IN
.
t
1
.
'^
'
DEC 14 1973,
DtWfcY LIBffAKY
WORKING
P>
.OAN SCHOOL OF
MANAGEMENT
PLANNING AND DESIGN OF SERVICE CONTRACTS
FOR CONSUMER DURABLE GOODS
Uday
S.
685-73
Karmarkar
November 1973
MASSACHUSETTS
TECHNOLOGY
50 MEMORIAL DRIVE
CAMBRIDGE, MASSACHUSETTS Oi
INSTITUTE OF
MAS!
INST.
!•
DEC 14 1973
DEWEY
PLANNING AND DESIGN OF SERVICE CONTRACTS
FOR CONSUMER DURABLE GOODS
Uday S. Karmarkar
685-73
November 1973
;
no, tRs*-?*>
7
1974
INTRODUCTION
:
The term "consumer durables"describes consumer goods that have the
characteristic of providing
a
constant stream of services over the
period of their useful life. Examples of such goods are household
appliances such as washing machines and refrigerators.
In goods of this
type questions of reliability, durability and maintainability of the
product become important, especially where the product is long-lived
and not subject to rapid obsolescence or vogue change.
The properties of reliability, durability etc. enter into the costs
incurred in providing continuous service over the life of the product.
These costs are due to repair, maintenance, servicing and part replacement. Clearly the greater the reliability of the product the less these
costs will be. However the improvement of these characteristics can
only be bought at a price
costs
-
-
that of increased design and production
which is going to be reflected in the price of the product
to the consumer.
From the buyer's point of view there exists therefore,
some optimal level of reliability and maintainability depending on
for present and future costs.
his relative preferences
As for the manufacturer
a
,
he is expected to provide a product that delivers
given stream of services at
the sale price.
a
level of reliability commensurate with
However, given that consumers need not be accurately
informed about what future costs to expect, there is no incentive for
manufacturers to be maximally productive with regard to product reliability.
That is to say, if we plot on
a
graph points representing technologically
feasible combinations of price versus future repair costs, it may be possible
for manufacturers to operate at points far away from the production-
possibility frontier which is formed by the southwest boundary of the set
of feasible combinations. (Figure 1)
.'1
iTO
2-
The economics of the situation, and the behavior of consumers and
manufacturers
have been analysed in detail by Hausman[l] and
Greenwood [2]. The interaction with prices in markets for used durable
goods has been extensively discussed by Dana [3]. The general area of
repair and maintainance in the household appliances industry has been
reviewed in [4] together with
a
discussion of significant problem areas
and policy recommendations. Suffice it to say that it is possible for
diseconomies to exist in the present system where responsibility for
repair and maintainance of the product lies entirely with the consumer.
SERVICE CONTRACTS ....
Hausman [1] and Greenwood [2] have suggested the mandatory offering of
service contracts by manufacturers as an elegant solution to the problems
of productivity in the area of repair and maintainance of durables. A
service contract is an agreement between the manufacturer and the buyer
by which the manufacturer undertakes to repair and maintain a product
for a fee payment by the owner.
Such an offering would leave the customer
the option of buying the contract or choosing to bear the costs of repair
etc.
himself.
In either case the result would be that the
consumer is
fully informed about what magnitude of repair costs he might expect to
incur.
Furthermore the system benefits from such an arrangement since
with the centralization of the repair operation in
a
factory environment,
economies of scale can be realized due to improved planning and effective
resource allocation and utilization.
In short
,
a
very good case can be
made for mandatory service contracts from the viewpoint of consumer pro-
tection and overall system economy.
...
and the MANUFACTURER
:
For the manufacturer however, there is no distinct advantage to be gained
from such an offering unless he feels that he has
a
competitive technolo-
gical advantage in the area of product reliability, that mandatory service
contracts would allow him to exploit.
Indeed several manufacturers do in
fact emphasise reliability in their marketing approach (Maytag, VW for
example) and some already offer
a
service contract option.
Given that
manufacturer has decided to offer
a
a
service contract, it
in his interest to ensure that he derives due benefit from the offer,
is
since he is contracting to provide services that would otherwise be provided
by some other profit making organization. The contract should be seen as
a
product to be appropriately priced and marketed and for which production
and inventory planning decisions have to be made. Some of the many decision
problems that might arise in planning and designing service contracts are:
i)
Design of the contract.
Capacity planning for repair and service facilities.
ii)
Inventory policies for replacement parts and repair resources.
iii)
iv) Optimal
v) Optimal
cost-reliability tradeoff.
product testing and quality control effort.
vi) Design of data collection and information systems.
vii) Special problems.
In making these decisions the
manufacturer should employ all the information
he can obtain about the failure characteristics of the product.
If it
possible to model and predict part failure, the relationship between
product cost and reliability, expected costs of repair etc., the decisions
involved can be made on
a
systematic basis.
DECISION PROBLEMS IN PLANNING AND DESIGN:
All
the decisions described below can be seen to be highly interrelated.
However the issue of interactions
is
deferred for the present for the sake
of expositional clarity.
1)
Design of the Service Contract
:
This relates to the design of the terms of service contract offer. The
main variables involved are
i)
:
Price
ii) Time horizon
iii) Extent of services to be provided.
Price
:
This is naturally related to the other terms of the contract.
Once those are fixed, the price must be set not only on the basis of profit
margin considerations and cost of providing the services contracted for,
but also on the marketing and promotional value of the contract and other
-4-
competitive factors. Thus the setting of the contract is a highly complex
decision involving many subjective factors.
However important inputs to
the decision are the cost of providing services, and the demand curve for
the service contract.
Time Horizon
Clearly, the contract cannot cover all servicing
:
to infinity and in general
the longer the time of coverage, the greater
the marginal cost to the manufacturer and the greater the price of the
contract.
It would probably be necessary
to cut off coverage before
failure rates of the product start to increase drastically, and the time
of coverage is thus also informative to the consumer. There are many other
factors that intervene; for example even if the product is discontinued,
service and spare parts have to be made available. From the point of
view of consumer perception and behavior, it may only be necessary to
ensure coverage for the first couple of years after purchase to provide
the buyer with enough security. This is especially true of durables which
are highly vogue dependent and have high rates of obsolescence. The key
inputs to this decision would be failure models, the costs of service for
different time horizons, models of consumer behavior in terms of original
buyer usage and total usage, etc.
Extent of Services
Coverage should presumably be sufficiently broad
:
to remove from the buyer any sense of insecurity or uncertainty;
i.e.
the
buyer should be protected from the onus of bearing large costs due to random
failures. However, the time of coverage might be set differently for different
parts and subsystems, and some items could conceivably be excluded from
coverage. Again failure rate models are important in making these decisions.
2)
Facilities Planning
:
Facilities supporting a service contract offering would include repair shops,
spare part and repair resource inventories, product collection and return
systems, manpower organization, data collection and information systems etc.
The capacities required are determined based on the anticipated volume of
requests for repair and service under the contract. These in turn depend
on the number of contracts in force and rates of part and subsystem failure.
We thus need models of demand for the contract as well as failure models.
3)
Inventory Policies
In
determining inventory policies, the demand for replacement parts and
:
material requirements can be predicted from knowledge of failure characteristics. Given the population of units in service the demand for
service can be predicted based on statistics from failure models such
as renewal
4)
or restoration processes.
Production Cost
The less reliable
Reliability Tradeoff
-
a
:
product, the greater the costs to the manufacturer of
providing service and repair.
If the price and time horizon of the service
contract are determined by considerations other than these costs, then
it is worthwhile to build greater reliability into the product.
on the appropriate level of effort in this area, a model
To decide
is required that
relates reliability to design and production costs.
5)
Product Testing and Quality Control:
Testing provides improved knowledge about the reliability of the product.
In general
better knowledge and reduced uncertainty reduce costs by allowing
planning to be done more precisely with less costs of safety margins and
hedging. To cite a simple example
—
better predictions of demand for
replacement parts reduce:safety stock requirements and hence the associated
holding costs. Similarly, improved quality control procedures would lead
to a reduction in service and repair requirements.
In formulating decision
problems in these areas, we need models that enable us to place a value
on information about failure rates.
6) Data Collection and Information Systems:
Incoming consumer requests for repair and service are informative in themselves and ought to be tracked. Here it is necessary to determine what
data is significant and how it is to be used in improving decision making.
7)
Special Problems
:
Unforeseen situations with large associated losses may suddenly occur as
in the case of the detection of a hazardous defect in the product.
cannot be explicitly planned for
a
These
except from the viewpoint of maintaining
tracking and detection facility included in the data collection system.
THE MODELLING APPROACH
:
The preceding discussion of decision problems has revealed a need for
models in the context of which decision problems can be framed. Here we
we discuss an approach to model building in terms of the basic models
that appear to be required
1)
Failure Rate Models
.
:
We require descriptive models of failure of parts in the form of probability
distributions of the time to failure (TTF) of each product element. There
may be dependencies between failure of different parts and joint distributions may be appropriate. Depending on the terms of the service contract,
we may need models of subsystem failure, "catastrophic" or final failure,
etc.
In addition by using Bayesian methods, we can incorporate the concept
of updating astate of knowledge as information becomes available.
2)
Sequences of failures
:
The occurrence of failures of a part followed by repair or replacement
gives rise to
a
sequence through time which can be modelled as
a
renewal
or a restoration process. Again a Bayesian treatment would allow us to
update our knowledge of the process as we observe it. Statistics from
these models can be used to predict aggregate demand for service.
3)
Sequences of costs
:
Each failure and service request gives rise to an associated cost of repair
or replacement [r/r) The present value of such a sequence of future costs is
an important input to various decisions.
The expected value of these cousts
discounted to the present is a very basic model in the present context.
4)
Production Cost/Reliability Model
One way to approach the problem to relating costs of production and design
to product reliability might be through the prior distributions for the
parameters of the time to failure (TTF) model. We could subjectively
estimate the prior conditional on the production costs; or we could assert
that increased production and design efforts lead to some change in the
prior parameters. Alternatively we might argue that over
a
small range,
the time to failure changes linearly with incremental change in design and
material costs, and use Bayesian regression models. However this type of
7-
model would greatly complicate
the analysis, and furthermore in the
interests of tractability we would have to restrict considerations to
Normal
5)
probability distributions.
Value of Sampling and Information
:
Bayesian models enable us to treat questions such as the value of information,
gain from sampling, and optimal sampling. These models are the basis for
tradeoffs involving costs of information gathering as in the cases of
testing, quality control and data collection that were described earlier.
The value of information is determined by comparing expected cost without
the information against expected cost under the updated state of knowledge
after obtaining the information.
6) Demand For Service Contracts
:
The demand for service contracts determines the overall capacity levels
required in planning for service and repair facilities. This demand is
a
function of the price of the contract, and consumer perceptions of
reliability and repair costs. It is clearly bounded from above by the
demand for the product itself. Regression models might be useful in
prediction of this demand based on variables such as price, advertising
and promotional efforts, rate of obsolescence etc.
SPECIFIC MODELS
In this section we describe some
modelling efforts in greater detail.
The basic models relate to part failure and sequences of failures with
their associated costs.
Time to Failure (TTF) Models
:
The basic descriptive model of part failure consists of a probability
distribution on the time to failure (TTF). The literature in this area
is extensive and a wide
variety of distributions has been used in modelling
failure characteristics.
(
Barlow and Proschan[5], Zelen [6]). The Expo-
nential, Erlang, Weibull and Normal distributions are commonly employed
failure models.
For our present purposes we are interested in Bayesian models of failure.
This in effect involves considering the parameters of the TTF distribution
as not known with certainty.
Prior distributions are assigned to the un-
known parameters and are updated as new information becomes available.
Models such as the exponential and Normal distributions are well suited
to Bayesian modification.
The Weibull distribution however, does not
possess sufficient statistics of fixed dimensionality, and hence is not
amenable to the usual form of natural conjugate analysis although Bayesian
treatments are available in the literature (Sol and [7], Bassin [8]).
The type of model chosen in a particular case will naturally depend on
the failure characteristics of the product in question.
However we are
here especially concerned with the early stages of product lifetime, and
this emphasis warrants some discussion in terms of its effects on model
selection. Most products are subject to "infant mortality"; i.e.
is a segment of the
there
population of all units produced that fails very
early, possibly due to some undetected defect.
In the
case of complex
products with many parts there may be many distinct part failures in
short time, or there may be persistent failure of
subsystem.
In
a
a
particular part or
other words the population of all products contains some
.raction of defectives or "lemons".
The effect of such a subpopulation of defectives on failures observed is
shown in Figure
2
which shows a record of cumulative service requests(in
terms of %of total sales) versus time measured from date of sale. The numbers are hypothetical but the general characteristics of the curve are
fairly typical.
cumulative
requests for
repair/service
{% of total
sales)
G
9
»z.
15"
2i
lg>
14,
Time from date of sale (months)
Figure
2
The initial transient portion of the curve suggests the existence of
defectives with a relatively high failure rate. As units are repaired
and returned to the population, the proportion of defectives and the
overall
failure rate both decrease. The asymptotic straight line portion
would thus represent
a
constant average failure rate which may be inter-
preted as being due to the"usual" incidence of random failures.
The Weibull process is often used to model such behavior exhibiting
decreasing failure rates. However, it is not an analytically tractable
model, especially in the Bayesian context, and estimation of its para-
meters is
a
cumbersome process
(
Hausman and Kamins [16]). In looking
for alternative models, it was hypothesised that
consisting of
a
a
failure distribution
mixture of exponential densities would exhibit the
desired failure rate characteristics and this can easily be verified.
Furthermore we can interpret this model intuitively as representing
mixture of two kinds of units --
a
proportion
high failure rate (A,) and a proportion '1-p'
'p'
a
of defectives with a
of "normal" units with
lower failure rate (A ) ascribed to random failures.
2
a
-10-
Sequences of Failures and Costs:
repair or replacement so that the
The sequence of failures followed by
unit is "as-good-as-new" constitutes
a
renewal
process. This type of
process has been extensively modelled and analysed in the literature
(Cox [9], Zelen [6]). We may also wish to consider restoration process
models where servicing does not leave the unit as good as new but rather
restores it to its state at the time of failure so that it is "as-bad-
as-old" (Ascher [10]). Suppose that failure of
a
unit can be caused by
one of a number of parts. Then failure of one part followed by its rep-
lacement does not render the unit as good as new and hence
a
restoration
model might be appropriate. Both types of models have been studied in
a
In the following however, we shall
Bayesian context (Bassin [11]).
chiefly be concerned with the renewal model.
Consider the stochastic failure of an element of a system. Suppose that
each failure is followed by (instantaneous) repair or replacement of the
element. Then the resulting sequence can be modelled as a renewal process
with associated costs of renewal. Many of the results that follow are
essentially available in Cox[9] but are repeated for continuity; the
notation we adopt paraphrases that used by Cox
t;.
:
time to failure of the i'th component.
:
ft*): p.d.f.
(of time to failure)
I£t|e): p.d.f. conditional on parameter set
<r'(9le>')
:
Prior distribution on
prior parameter set
1
D(t[^ ): p.d.f. of
£
&
in the Bayesian sense, with
_£*
unconditional on the uncertain parameter
set
F(t)
G ^)
2C±)
:
=
•
c.d.f. (of time to failure)
jf Ki)<K
PLt»t3-i-Fft)
-
(Instantaneous) Failure rate or Hazard rate function.
= d{$('); $\
<fo)
?<*>- Etc"**] -
s
•
The Laplace transform of
±{ K-);s]
a
function
$0)
Total time to the r'th renewal
S-,
S,
- t, +
t4+
..
+ tn
.
fe^t) :.p.d.f. of
?,
K
S,
c.d.f.
Y (t)=
fj fc
H(t)
of
:
The number of renewals in (0,t)
•
The renewal
-UCt) = i. H(t)
of
L(»
=
E(N t2
)
function:- the expected no. of renewals in(0,t)
the renewal density
:
1
V
H (N t
)
s
L(t)-
LHCt)]"*"
With regard to the sequence of costs associated with renewals, we are
interested in the expectation of future costs incurred, discounted to
the present. There are several different quantities that are of interest.
We may want to know the cost of an infinite sequence of renewals or of
a finite number of renewals.
In the
service contract context in particu-
lar, we might want to know the discounted costs associated with a sequence
of renewals that is truncated at
the number of renewals.
R(t)
:
a
finite time, without restriction on
Let us denote
:
the cost associated with a renewal at time
t.
assume for the present that this is constant,
^
:
We will
R£t)=
R
discounting factor for continuous discounting of future
costs.
"V- Y
V-<
V
:
discounted cost of the r'th renewal;
:
total discounted cost of the first r renewals
:
TDC of an infinite sequence of renewals
-
\y(
T)
:
Z^
TDC of a sequence of renewals truncated at time
T (but
with no limit on the number of renewals).
As usual
the tilde is used to identify random variables (r.v.'s).
A bar will
-12-
be used to denote expectation of a r.v.:
discounted cost of the r'th renewal,
-vY represents the expected
V
is the EDC of a sequence
truncated at time T and so on. We shall assume that the TTF's of
successive elements are mutually independent, identically distributed
(m.i.i.d.) r.v.'s, conditional on the parameters of the distribution
being known. We first derive general formulae for the quantities of
interest, and then apply them to particular failure distributions.
CASE
I
:
Parameters of the TTF density known with certainty
For a single failure followed by renewal
v-
rt^s
:
0)
R i*tf)
,
/
Since the
is equal
are m.i.i.d.
r.v.'s the transform of the sum of the
to the product of the transforms, and hence
^
=
fcjfc* Cf )
=
X
X
'
c
s
:
i*(?)
R.TT
For a sequence of n renewals, the total expected discounted cost
:
and hence for an infinite sequence of renewals,
i-P<»
We note that since for the usual values of f
_ ft
t
e
$(*"
)
< ££fc)
,
we have
13-
To find the variance of this quantity first note that
For
j>(.
we can write Sj * S; + 1^ +
,
+
..
.
+1
term in the equation above can be replaced by
and hence the second
•
/
X
:
2. '2-
£
<~
w
"'
Whence applying the expectation operator
which upon simplification gives
:
RM * (in
E(v-).
\
-pr?
)
J
and the variance can be found in the usual way.
If we now consider the sequence of renewals truncated at time T, then
1
O
Z.\
which can also be written as
fT)
V
fcjV
,
ft
Jk(t-)
CO
where
J^(
t
\
On
=it
,
"S"
U
/(-)
we could easily replace
^
=
°L m^t
)
"»
the renewal density. Clearly
with some time dependent function
R-Ct).
Again in determining the variance of this quantity, since by truncation
we are saying that the cost of
"
a
renewal
for
Ul
t>T
I-
I
is zero we can write:
1
-
«-+
'
g
which on simplification yields
fltM'l.
L
J
where
:
"
ft*[i-Kn1, fY
[»
-l(f)l
i
2<?fr
A(t)jr
rt)
Considering now the number of renewals in (0,t)
N t <.T implies
We know that
and hence we can write
the renewal
function
Therefore
P[w t =>]r K y tt)- k
£
(Cox[9])
:
^[N t <v]=
X y Lt)
-
(t) -We can thus determine
K
(t
=
^]
£^I> v tO-Kv
+
=
^)]
(»o)
)
:
LIAVO
s
-fe*Y
(O-
[ £*(*•)]
sli
-
hence
rn)
iWD
f>no]
HCt)=
In finding the variance of the number of renewals,
=
I
*
and we know that
L(t)
P [S-r> t]--
^
£^P[Nt
-.
Taking transforms on both sides
H ^0
.
:
Hit), ECN t)
WCt)--
S^>t
E(V)
=
Z^
V
Taking transforms
,
V
first find
[^vCt)-W Y+ ,:'t)]
s
simplifying and noting that
"*
ZL
v:o
P
~
P
It
+ ^
^
- P
-
3
)
we get
)-
and we can thus find the variance as
Van t5 t ).
uo- |>>r
;
uo.i-'^'l
-15-
The distribution of the number of renewals may also be obtained by
studied the generating function of
Nt
(
Cox[9]). The results are
cumbersome, and we will not generally obtain this distribution. However, we briefly indicate the procedure involved.
v:
Taking transforms on both sides
-
And we can find
CASE II
:
i
PlN^J
-
Let us denote
\
:
j
»(o
as the coefficient of
£
in
H£t,
2 )
Parameters of the TTF distribution not known with certainty
One approach is to assign a prior distribution* (©.
I
fi'
) to
6_
the set
of parameters not known with certainty, and then to derive the unconditional distribution of time to failure for a single failure.
®
so that
—
v
=
(
RD*«
)
(it)
For a sequence of failures, the unconditional distributions of the
J
.otal
times to failure for successive failures are not independent, since
they share the same prior distribution on the uncertain parameter set.
Hence the simple results obtained in the previous case do not apply when
we use this direct approach:
16-
VT
£
R
^
=
i
D?(f)
08)
i
However a simple alternative approach is just to take the prior expectations
of the expressions obtained conditional on the parameter being known (Case
I).
Thus for example:
({
And similarly
}«e-**,(t|g)it} rCfi|i»J)Je
o
(9
:
1
vn)
s
E
V
[v
*vf,n
I
(t;
(q)
f22)
]
When considering variances of quantities, we have to use the appropriate
formula for iterated variances.
Applications to particular TTF models
(A) The EXPONENTIAL TTF distribution:
"
At-
\
At
-17-
For the exponential density:
st-
~ xfc
-
and hence direct substitution in equations (1),(3),(4) and (5) gives:
RX
=
-v,
&>)
v;
*Mi
v,.
From equation (7) we also have
and hence the variance of the total discounted cost of an infinite
sequence of renewals
~
"'"
V
To compute the quantity
»
^( t
)
we first find the renewal density
And by equation (8)
,. L >i.|
p.
,
£*-,(*;
-
Tr
=
^
(>
-D
:
-At
i. a
:
o
We can find the variance of this cost by using equation (9) which gives
F_(v^
HA(^-e" 2?T
2
]
=
2-r
where
:
I(r\ B
A
(
i
—
1
sT'*"
)
y
['
+*<«)
Li -i/r
)
1
J
c
*9)
f)" " \
1
The variance can now be found in the usual way.
In
determining the renewal function, we note that we have already obtained
the renewal density k(t
To find the variance of
Therefore
=
X
Nt
,
)
,
and hence
equation (12) gives
^
Suppose now that the parameter
with certainty, and we
is not known
assign to it a Gamma-1 natural conjugate prior density .(Raiffa &
Schlaifer [12]):
_Xt'
,i
*,V,t'>0
,
OC
-fyiC^W.t')
5\
*
*.
;
(32.)
The unconditional density of the total time to the r'th renewal is
given by
:
Dv OKt';v)= Jf w Cth?0X(*Kt
,
W>
C33)
o
and has been shown in [12](7.4.2) to be of the"inverted Beta-2" form:
^CtKt'
;
v)= f . Ct|T
t a
v t')-
f
#
_j
t'
BO, V)
0+fJ y+Y
We could presumably obtain directly quantities such as
'
t'
:
o
However this "direct approach" is explored in Appendix A and is found
to be quite cumbersome.
manner in equations (25)
X
Noting that the parameter
-
(31) we can simply write
appears in a simple
:
M£)
v
s
where
>'= j'
T
is the prior
expectation of X
.
Other results follow
1
in a similar manner.
For example
:
*(££) +
^«*(>/)
=
^N'
f
V
->(V)
z+- -ft')
r3?
'
-19-
(B) The MIXED EXPONENTIAL TTF distribution:
This distribution was suggested as a model adequate to describe failures
exhibiting
a
decreasing failure rate. As stated earlier, the intuitive
notion motivating this model conceives of the population of all parts as
being composed of two sub-populations
fraction p subject to both
a
:
defective and random failure, and a fraction (1-p) subject only to
If it is assumed that both sub-populations exhibit expo-
random failure.
nential
A
failure distributions with rates
may write
and
,
>\
x
respectively, we
:
A,t
F4,(t) +(>-p)j,(t)
Jft) =
We may generally assume that
=
pA,c"
p<
*,>*2. and
f
»
*0-P)A x
e"
x
- p.)
We will first examine the failure rate characteristics of this distribution
At
t-.o
we see that
££o)r
p
x,+
6-
p
)>
too)
j_
Furthermore, Z(t) is monotone decreasing in t since considering the sign of
the derivative of Z(t)
^-2Ct)
:
- CA
=
[
.-Q
Pt
'
-f\ + A x )t
f
Q-P)
<
e
+ (>-p)c
(41
a
J
Rewriting this as
2L2Ct
at
)
=
-(\-^'? (| - f)
i P 0-p)< p e -K-Ot
We can see that for any
=-
+(,.piV
>
(-
"- x >t
'
X,, Xj.
(4»)
and we can also see by writing
)
)
20-
P A '*
Ut)
that for
+ Q-f)*a.
(«3)
>, > >2_
f4<0
t -*<»
We define here for future convenience the quantity
z.
4<n
D
t=o
which might be interpreted as
a
sort of measure of the heterogeneity of
the population being modelled.
We have thus shown in equations (40) through (43) that the mixed exponential failure density exhibits a decreasing failure rate
Figure
:
3
Let us now consider the behavior of the renewal function for this density:
res)
=
_L±L
*
L'-P)>-
'M)
Using equation (11
whence after some algebra we obtain
W\ S )
=
>>,. +
sl>V
(Up
)Xj
(Ml)
-21-
Defining
-A,
P^
=
7^
.
+(l
"P )X i
(i-p)A,
+
P X,
(A 8)
=
(>,
*M~-A-,
We can rewrite
To take the inverse transform, we first note that
rso)
and hence upon simplification we can write the renewal
function as
Investigating the coefficient of the exponential decay term
:
The renewal density is given by the slope of the renewal function:
(**)
and we can see that
-k(o)
--
V\_
v
=
2^)
(54)
The shape of the renewal function thus looks yery much like the curve
in Figure 2
Jiis model
in its general
will
behavior, and is shown in Figure
in fact fit actual
4.
data remains to be verified.
Whether
-22-
H(t)
Figure
In computing the
4
variance of the number of renewals, substituting the
appropriate quantities in equation (12) we get after some algebra that
^
L
"
(7~>
{
A
+
l^M^i)
(SS)
]
To write the inverse transform let
fsM
and note that
•A x t
a 2 tt)
^
t«£
=
A xt
A,
i-
l*tol-
£-
V *^*^
1
t")
'
23-
whence we can write
L(t)
=
Z_ X.
I
=
(58)
«-'Jt)
\
Examining now the cost models corresponding to this failure density, we
can write from equations (5) and (8) that
We will not write out the variances of these quantities in all detail.
We remark however that in equation (7) the multiplying factor can be
written as
|
LLC
\
-
ret)
and in equation (9) we can write for the integral
OXOt
Parameters not known with certainty
i
f
a
r,
-
Cf-tX^T
:
Estimation of the parameters for the mixed exponential density is
trivial
a
non-
task. While this distribution possesses desirable characteristics
and gives a more intuitive interpretation of the failure process than
say the Wei bull distribution, it does not buy us tractability in a Bayesian
approach. The distribution does not possess sufficient statistics of
fixed dimensionality and hence cannot be assigned a natural conjugate
prdor in the usual way. However, Behboodian [17] has investigated the
Bayesian treatment of the mixing parameter
p and
shows that the use
of a Beta prior on p leads to a mixed Beta posterior density. Thus the
r
amily of mixed Beta densities is conjugate to the likelihood function
for the mixing parameter p of a mixture of distributions. The number of
terms in the posterior mixture depends on the sample size.
-24-
The estimation problem in
in [18] and [19].
a
classical setting has been studied by Boes
Rider [20] has examined the application of the method
of moments and Mendenhall and Hader [21] have studied the maximum like-
lihood approach to the mixed exponential case.
In general
we may obtain information about the process in different ways.
For example the data could be from
i)
:
A random sample that is subjected to a life test for a fixed period
at the end of which there may or may not be survivors,
ii) Historical data on the aggregate rate of demand for service requests;
i.e.
cumulative failures as in figure
2.
iii) Presently incoming data on service requests, where it may be possible
to obtain some additional
information on the failure.
The treatment of the problem varies depending on the particular situation,
and the assumptions made about the sampling process.
For example if we
assume that a failure can always be identified as having occurred from
one or the other of the subpopulations, the problem is considerably
simplified and becomes
a
routine case of prior to posterior analysis. This
assumption may not however be very realistic.
In particular we note that
for the interpretation suggested here, it is possible for a defective
element to fail in
a
random mode and be indistiguishable from
a
random
failure although the reverse might not be the case.
The practical
importance of estimation of the parameters is directly
related to the decisions that are to be made.
It may be that
A
2.
the
failure rate for random failures is quite well known from experience
with similar products.
In any case it is
probably not a controllable
variable. The parameters A, and p on the other hand might be considered
control variables in the following sense
:
If more effort is put into
testing screening and quality control one could hope to reduce the
fraction of defective items in the population. Similarly the control of
>
1
might be achieved as a result of deliberate manufacturing policy
interms of reliability and durability designed into the product. Furthermore testing can improve knowledge about these parameters and this will
affect the decisions that are made. Thus we see that these parameters have
a
natural
interpretation in
a
decision making context.
-25-
MODIFICATIONS
&
EXTENSIONS
We have outlined the basic approach to modelling sequences of failures
and costs with a view to formulating decisions about planning and
designing service contract offerings.
It is relatively simple to make
certain modifications to these models to fit particular situations.
Some of the many possibilities are briefly sketched out below
a) Service requests and costs for a group of identical
:
products
:
Assuming that these fail independently, the expected number of requests
for service is given by the sum of the expectations.
hold for the variance of the number of requests
,
Similar results
expected costs and
variance of the costs.
b) Different dates of sale
:
If the sale dates are known for two batches say, the expected number of
requests can be found simply by adding graphs such as in figure 4, suitably
translated along the time axis to account for the date of sale. If the
sale date is a random variable, then we can use the artifice of regarding
sale of the product as the first "renewal". The process can now be
modelled as
a
"modified" renewal process (Cox [9]
in a sense a control
).
The date of sale is
variable, since by staggering sales in time, it is
possible to smooth out the demand for service requests.
c)
Different failure densities for original and replacement parts
Suppose we can assume that
repaired product has
a
a
:
different failure
density from the original -- for example we might assume that there
are no defectives amongst replacement parts. Again, a modified renewal
process would be an appropriate model.
d)
Different cost functions
:
We can see from equations (8) and (28) that it is fairly straightforward
to consider other than constant costs.
In particular we may want to use
polynomials or exponential functions in t to reflect costs which change
systematically over time. We may also wish to consider stochastic costs
ind utilities.
Apart from these modifications, it may be of interest to determine other
information from the same basic model. For example, the average age of
parts in service can be determined by finding the mean "backward re-
currence time
".
-26-
APPENDIX A
:
In this appendix we explore in some detail
the approach suggested in
equation (35). That is to say we attempt to obtain some expressions
for future discounted costs for the case of exponential TTF distri-
butions, by working directly with the unconditional sampling density
obtained by assigning Gamma-1 priors. The failure density unconditional
on the parameter
equation (35)
A
is of the "inverted Beta 2" form.
Recalling
:
Let us denote the integral:
^
-S x
/
"»-\
/
i
(X
O
-V
o
')~'
T
,
so that
For integral values of r and r'
Erdelyi et.al.[14] give that
Now r is integer but in general
r'
is not.
In this case we obtain from
Gradshteyn and Ryzhik [13] that
c
!(*,*•
where
W
0)is
o
=
l
;l
^'i
c^ITo
W,.*v->'
a.'^i>)
;
a.v
the Whittaker Function which is written as
where
The confluent hypergeometric function ,F,
,F
io/
V
? ^
-
I
4-
* *
Substituting these expressions in A
+
A
denotes the series
±j£±0
we have
*%
^(c,tl
)^+;)Z
'
+
..
U,7)
:
^f'V.c*,.-*'; *,)>*]
i* .;
;
-27-
For the case r=l we can directly evalute the integral
v;
Ue
.-
st
ill
-J
&ov
i
(t
)
AC
,
4 t')
i+y
T(-
which can be found from the handbooks to be
1
(:/^"
-
iif/^'Etf-ft
1
ca.»o3
;
where the exponential integral is denoted by
-Ed-It')
-
E,(fi'>
t«"V
J
^
U)
^
ft'
Clearly evaluation of
V
say by direct summation of costs is going to
be quite involved due to the extreme unwieldiness of these expressions.
V
Of course we have already determined
in
equation (36). We shall
here demonstrate that we can reproduce these simple results by directly
working with the inverted-Beta-2 density. Let us write
Co
<x>
I
Reversing the order of summation and integration and simplifying
where we have again set a
=
for notational convenience.
t'
Now this interchange is feasible if and only if the series in brackets is
uniformly convergent in the the sense that the sequence of functions formed
by the partial
sums of the series converges uniformly (Rudin [15]}.
In
fact if we proceed as above, this is not the case. To see this consider
the sum
and writing k =(r-l)
S(*V
j
2
we have
)
-
r
,
^»
^
T his is therefore
a negative binomial
n
Now as >- >oo
j
;
?
->
1
,
,
rfk+ of(V+o
series and hence
,-(A.)
VM
and so the series diverges. However if we exploit
-29-
the terms in the integral outside the summation sign, we can circumvent
this problem. We consider the expression
Now since *
t
C"<vt]
W e have ?<
f
.
TT e,]
and nence the series does not
diverge. Thus replacing
we have
/
and recalling that we had set a = t
7
-
*~
v
(T)
i
-,
1
=
R
^
which fortunately agree with the results obtained earlier, since
£"^f'M
-
\_
.
-30-
REFERENCES
[]]
:
Hausman, W.H.; "Improving Resource Allocation, Productivity and
Consumer Information by Mandatory Offering of Prepaid Service
Contracts"; Internal Memorandum (unpublished); M.I.T.(Feb 1972).
[2]
Greenwood, J. M.
;
"Service Contracts as an Incentive to Product Durability
Reliability and Maintainability"; Thesis, Mgmt. (1972) M.S.;M.I.T.
[3]
Dana, J.
D.
"Some Aspects concerning the Theory of Consumer Durables";
;
Thesis, Econ.(1967) Ph.D.; M.I.T.
[4]
Report of the Task Force on Appliance Warranties and Service; U.S.
[5]
Barlow, R.E.; Proschan.F.
[6]
Zelen, M.(ed.); Statistical Theory of Reliability
Federal Trade Comm. & other agencies;
J.Wiley & Sons;
(1969).
Mathematical Theory of Reliability ;
;
N. Y.
(1965).
;
U.of Wise. Press;
Madison (1963).
[7]
Soland,
R.
;
"Bayesian Analysis of the Weibull Process with Unknown
Scale and Shape Parameters"; IEEE Trans. Re!
.
;R -18 No.4(Nov.l969).
[8]
Bassin, W.M.; "Increasing Hazard Functions and Overhaul Policy";
[9]
Cox, D.R.; Renewal Theory
[10]
Ascher, H.E. ;"Evaluation of Repairable System Reliability using the
[11]
Bassin,W.M.; "A Bayesian Optimal Overhaul
[12]
Raiffa, H. ;Schlaifer,
IEEE, Proc.
1969 Ann. Symp.Rel.
;
;
Chicago (1969).
Methuen, London (1962).
'Bad-as-old' Concept";
IEEE Trans.
Rel
;
R-17 No.2;(June 1968)
Interval Model for the
Weibull Restoration Process"; to be published, (1972).
R.
;
Applied Statistical Decisi-on Theory
Colonial Press; Clinton Mass.
Tables of Integrals, Series & Products
[13]
Gradshteyn, Ryzhik
[14]
Erdelyi, Magnus et.al.; Tables of Integral Transforms, Vol
;
;
(1961).
.
I
;
McGraw-Hill Co.; N.Y. (1954).
[15]
Rudin W.; Principles of Mathematical Analysis
;
McGraw-Hill Co.;
N.Y. (1964).
[16]
Hausman W.H.; M.Kamins
;
"The reliability of New Automobile
Parts "; Annals of Maintainability and Reliability.
[17]
Behboodian, Javad; "Bayesian estimation for the Proportions in
a
Mixture of Distributions"; Sankhya; SeriesB, (1972).
-31-
[18] Boes, D.C.
"On the Estimation of Mixing Distributions";
Ann. Math. Stat
;
37, 177-188.
[19] Boes, D.C.
;
"
Minimax Unbiased Estimator of Mixing Distribution
for Finite Mixtures"; Sankhya
[20] Rider, Paul;
"
:
Series A; 29
The method of moments applied to
a
exponential deistributions"; Ann. Math. Stat.
[21] Mendenhall, W., Hader, R.J.;
,
417-420.
mixture of two
,
32,(1961).
"Estimation of parameters of mixed
exponentially distributed failure time distributions from
censored life test data"; Biometrika , 45(1958).
.
Download